SNAPKITTYWEST's picture
add README.md
1867bb7 verified
|
Raw
History Blame Contribute Delete
2.33 kB

Sovereign XML Compiler

Natural language → valid XML system prompts in one shot. Zero correction iterations.

Three modes using logit gating at the tokenization layer:

Modes

1. GBNF Constrained Decoding (best)

Uses llama.cpp grammar-based sampling. The model physically cannot output invalid XML. Token selection at the softmax layer is masked to enforce the grammar. Result: 100% valid XML, zero syntax errors, one pass.

# Requires llama.cpp server running
export LLAMA_URL=http://localhost:8080
python server/compiler.py --mode gbnf --input "You are a Lean 4 proof verifier..."

2. Skeleton In-Filling

LLM only fills {{PLACEHOLDER}} values. Parser injects into template. The model never writes < or > — no tag hallucination possible.

python server/compiler.py --mode skeleton --input "You are a sovereign math agent..."

3. Dual-Pass Chain-of-XML

<thought_process> first, <xml_output> second. Forces attention to compute correct structure before emitting markup.

python server/compiler.py --mode dual-pass --input "..."

Quick Start

export OLLAMA_URL=http://localhost:11434
export XML_MODEL=nemotron

python server/compiler.py --mode skeleton \
  --input "You are a zero-sorry Lean 4 verifier. Reject any proof containing sorry." \
  --output my_prompt.xml

Files

File Purpose
grammars/sovereign_prompt.gbnf GBNF grammar for llama.cpp constrained decoding
skeletons/sovereign_prompt.xml XML skeleton with {{PLACEHOLDER}} tokens
server/compiler.py Three-mode compiler server

The Gate Taxonomy

Mode Gate Type Mechanism
GBNF Weight-level structural gate Token masking at softmax — invalid tokens → probability 0
Skeleton Context gate LLM sees only leaf values, never writes tags
Dual-pass Attention gate CoT forces structure before syntax

Connection to Gates Normalization

GBNF masking is P(token | grammar) = 0 for forbidden tokens. This is the same logit gate formalism from the Gates Normalization paper: G_P(D_M) = softmax(logits_M + b_P) where b_P = -∞ for grammar-violating tokens. The simplex constraint holds: ∑P = 1 over the valid token set.

License

Sovereign Source License v2.0